決策樹與區別分析模擬測繪木荷之空間分布型態

Modeling and Mapping the Spatial Distribution Pattern of Schima superba Via Decision Tree and Discriminant Analysis

木荷（Schima superba var. superba）是常綠喬木，為優良闊葉樹種之一，廣泛分布於全台灣中低海300-2, 300m。木荷具有防火樹種特性亦為造林樹種之一，且GIS多用於珍稀物種，而少用於廣泛分布樹種，故選定為探討之對象。本研究係利用GIS疊合GPS定位木荷圖層與DEM導出海拔、坡度、坡向、坡面位置加上SPOT-5衛星影像導出植生指標等五項預測變數圖層，結合多變量統計決策樹（Decision Tree, DT）及區別分析（Discriminant Analysis, DA）模擬並測繪木荷潛在生育地之空間分布，並根據推測準確度與執行效率評比兩者的預測能力。研究結果顯示DT模式推測準確度明顯高於DA模式；在建模和驗模執行效率上，兩者難分軒輊，惟在繪製全區木荷潛在生育地，DT較DA略顯繁瑣。更重要的是，模式於初期模擬即由全區篩選出4-7%之高潛力區，大幅縮小後續地面調查之面積，從而節省龐大的經費與人力。因此，這兩模式皆可應用在木荷潛在生育地之模擬，惟DT法遠優於DA法。SPOT-5影像導出植生指標無法提升模式預測能力，乃因其空間與光譜解像力不夠細緻，無法分辨複雜而散生之林相。後續將融合高空間解像力、高光譜影像及光達資料，期能提昇模式預測能力。目前木荷樣株都來自東峰溪流域，難免空間自相關或樣本不足之顧慮，後續將增添關刀溪流域樣本，以再確認模式之可靠性。 Chinese guger-tree (Schima superba var. superba, CGT) is one of the fine broad-leaf tree species, widely distributed in central Taiwan with elevation ranging from 300-2,300 m. Chinese guger-tree has good fitment, with excellent fire resistance characteristics, and thus is one of major species of forestation. Many studies have applied a geographic information system (GIS) with statistical methods to model the habitat of the rare species, but not for the widely distributed tree species. Therefore, the species was chosen as a target for this study. GIS technique was applied to overlay the tree samples collected with GPS on the layers of elevation, slope, aspect, terrain position, and vegetation indices derived from SPOT-5 images for modeling the tree's habitat. Decision tree (DT) and discriminant analysis (DA) models were developed to predict and map the tree's potential sites in the Huisun study area, and to determine the optimum in terms of accuracy and efficiency. The accuracy of the DT model was much higher than that of the DA model; both were highly efficient in model development and validation, but DA was more convenient than DT for mapping the CGT potential sites over the study area. More importantly, the two models can be applied to model the tree's potential habitat because they greatly reduced the area of field survey to 4-7% of the entire study area at the first stage, but DT was much better than DA. Vegetation indices derived from two-date SPOT-5 images could not improve the predicting ability of models for the widely distributed species because of SPOT imagery lacking fine spectral resolution and spatial resolution. Fusion of high spatial resolution imagery, hyperspectral data and LIDAR data will be used in follow-up studies to improve the predicting ability of models. The conclusions remain further confirmed with independent samples chosen from the Guan-Dau watershed far from the Tong-Feng watershed where currently used samples were chosen due to considerations of spatial autocorrelation and insufficiency of tree samples.

Chinese guger-tree (Schima superba var. superba, CGT) is one of the fine broad-leaf tree species, widely distributed in central Taiwan with elevation ranging from 300-2,300 m. Chinese guger-tree has good fitment, with excellent fire resistance characteristics, and thus is one of major species of forestation. Many studies have applied a geographic information system (GIS) with statistical methods to model the habitat of the rare species, but not for the widely distributed tree species. Therefore, the species was chosen as a target for this study. GIS technique was applied to overlay the tree samples collected with GPS on the layers of elevation, slope, aspect, terrain position, and vegetation indices derived from SPOT-5 images for modeling the tree's habitat. Decision tree (DT) and discriminant analysis (DA) models were developed to predict and map the tree's potential sites in the Huisun study area, and to determine the optimum in terms of accuracy and efficiency. The accuracy of the DT model was much higher than that of the DA model; both were highly efficient in model development and validation, but DA was more convenient than DT for mapping the CGT potential sites over the study area. More importantly, the two models can be applied to model the tree's potential habitat because they greatly reduced the area of field survey to 4-7% of the entire study area at the first stage, but DT was much better than DA. Vegetation indices derived from two-date SPOT-5 images could not improve the predicting ability of models for the widely distributed species because of SPOT imagery lacking fine spectral resolution and spatial resolution. Fusion of high spatial resolution imagery, hyperspectral data and LIDAR data will be used in follow-up studies to improve the predicting ability of models. The conclusions remain further confirmed with independent samples chosen from the Guan-Dau watershed far from the Tong-Feng watershed where currently used samples were chosen due to considerations of spatial autocorrelation and insufficiency of tree samples.